Bringing graphics, X86 closer
"In 2011 we will deliver multiple APUs targeting multiple markets and our competitor will only have one targeting a specific spot in the market," said Moore, referring to archrival Intel. "And we are delivering a much clearer vision of architectural integration for graphics and compute cores," he said.
Moore outlined AMD's plan to evangelize the industry to help develop the software needed for merged x86, graphics chips. The stack includes new runtime environments, domain libraries and high-level frameworks.
AMD aims to make graphics equal citizens with x86 cores. The two will share unified virtual address space that is part of page-able system memory. GPUs on and off the main processor will be brought into the CPUs cache coherent memory domain, and ultimately GPUs will support context switching, Moore said.
Many of the features require support from partners such as Microsoft and developers of the OpenCL standard for multicore programming. For example AMD plans to deliver a GOU with a hardware scheduler with close OS integration.
"Any app in user mode will be able to use the GPU without talking to the OS or device drivers, making a direct connection to offload work with new task-parallel queuing runtimes," said Moore.
The future runtimes could be upgrades of today's OpenCL and Microsoft DirectCompute runtimes or new software such as ConcRT, GCD or TBB, he said.
AMD will also take some steps on its own. For example, it will support bi-directional movement of available power between CPU and GPU cores. It also plans a higher bandwidth on-chip memory controller.
AMD wants to develop a software stack for its Fusion chips.